US11943093B1 - Network connection recovery after virtual machine transition in an on-demand network code execution system - Google Patents

Abstract

Systems and methods are described for enabling graceful recovery of network connections in a virtual machine instance that has been migrated or temporarily halted. To prevent the virtual machine instance from attempting to reuse open connections, which might fail due to migration or halting, a host device identifies open connections just prior to halting the virtual machine instance on the host, and transmits to the virtual machine instance termination signals for the open connections. The host device may further transmit termination signals to the other parties to such connections. Each termination signal may be formatted so as to appear to originate from the other party to the connection, causing both parties to synchronize their knowledge of the connection. On reactivation, the virtual machine instance can recovery the connections without errors associated with attempted utilization of broken connections.

Description

BACKGROUND

Computing devices can utilize communication networks to exchange data. Companies and organizations operate computer networks that interconnect a number of computing devices to support operations or to provide services to third parties. The computing systems can be located in a single geographic location or located in multiple, distinct geographic locations (e.g., interconnected via private or public communication networks). Specifically, data centers or data processing centers, herein generally referred to as a “data center,” may include a number of interconnected computing systems to provide computing resources to users of the data center. The data centers may be private data centers operated on behalf of an organization or public data centers operated on behalf, or for the benefit of, the general public.

To facilitate increased utilization of data center resources, virtualization technologies allow a single physical computing device to host one or more instances of virtual machines that appear and operate as independent computing devices to users of a data center. With virtualization, the single physical computing device can create, maintain, delete, or otherwise manage virtual machines in a dynamic manner. In turn, users can request computer resources from a data center, including single computing devices or a configuration of networked computing devices, and be provided with varying numbers of virtual machine resources.

In some scenarios, virtual machine instances may be configured according to a number of virtual machine instance types to provide specific functionality. For example, various computing devices may be associated with different combinations of operating systems or operating system configurations, virtualized hardware resources and software applications to enable a computing device to provide different desired functionalities, or to provide similar functionalities more efficiently. These virtual machine instance type configurations are often contained within a device image, which includes static data containing the software (e.g., the OS and applications together with their configuration and data files, etc.) that the virtual machine will run once started. The device image is typically stored on the disk used to create or initialize the instance. Thus, a computing device may process the device image in order to implement the desired software configuration.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG.1 is a block diagram depicting an illustrative environment in which an on-demand code execution system can operate to execute tasks corresponding to code, which may be submitted by users of the on-demand code execution system, and to transition execution environments into various memory states based on an expected timing of a next execution of such code;

FIG.2 depicts a general architecture of a computing device providing a worker manager on the on-demand code execution system ofFIG.1, which may function to manage a memory state of an execution environment based on an expected timing of a next execution of code provisioned within the execution environment;

FIG.3 is a flow diagram depicting illustrative interactions for determining a memory state in which an execution environment should be placed based on an expected timing of a next execution of code provisioned within the execution environment of the on-demand code execution system ofFIG.1;

FIG.4 is a flow diagram depicting illustrative interactions for altering the memory state of an execution environment to reduce computing resource usage of the on-demand code execution system ofFIG.1 while maintaining the ability of the system to rapidly execution of code within the environment;

FIG.5 is a flow chart depicting an illustrative routine for modifying a memory state of an execution environment within the on-demand code execution system ofFIG.1 based on an expected timing of a next execution of code provisioned within the execution environment;

FIG.6A is a flow diagram depicting illustrative interactions for ending an external side of stateful connections of a virtual machine instance on an on-demand code execution system to prepare for migration of that virtual machine instance;

FIG.6B is a flow diagram depicting illustrative interactions for ending an internal side of stateful connections of a virtual machine instance on an on-demand code execution system to prepare for activation of that virtual machine instance, and for reestablishment of stateful connections after activation; and

FIG.7 is a flow chart depicting an illustrative routine for ending stateful connections of a virtual machine instance in connection with transitioning of the virtual machine instance, to facilitate graceful recovery of such network connections.

DETAILED DESCRIPTION

Generally described, aspects of the present disclosure relate to an on-demand code execution system enabling rapid execution of code, which may be supplied by users of the on-demand code execution system. An on-demand code execution system may also be known as a “serverless” execution system or a request-drive code execution system. More specifically, embodiments of the present disclosure relate to facilitating recovery of stateful network connections of an execution environment on the on-demand code execution system when those connections are expected to be disrupted, such as by migration of the execution environment to another device or by a change in state of the execution environment (e.g., suspension or deactivation).

As described in detail herein, the on-demand code execution system may provide a network-accessible service enabling users to submit or designate computer-executable code to be executed by virtual machine instances on the on-demand code execution system. Each set of code on the on-demand code execution system may define a “task,” and implement specific functionality corresponding to that task when executed on a virtual machine instance of the on-demand code execution system. Individual implementations of the task on the on-demand code execution system may be referred to as an “execution” of the task (or a “task execution”). The on-demand code execution system can further enable users to trigger execution of a task based on a variety of potential events, such as detecting new data at a network-based storage system, transmission of an application programming interface (“API”) call to the on-demand code execution system, or transmission of a specially formatted hypertext transport protocol (“HTTP”) packet to the on-demand code execution system. Thus, users may utilize the on-demand code execution system to execute any specified executable code “on-demand,” without requiring configuration or maintenance of the underlying hardware or infrastructure on which the code is executed. Further, the on-demand code execution system may be configured to execute tasks in a rapid manner (e.g., in under 100 milliseconds [ms]), thus enabling execution of tasks in “real-time” (e.g., with little or no perceptible delay to an end user).

The on-demand code execution system may implement a variety of technologies to enable rapid execution of code. Illustratively, the on-demand code execution system may be configured to maintain a number of execution environments, such as virtual machine instances, software containers, or the like, in which code of a task may be provisioned and executed. In some instances, an appropriate execution environment may be currently executing on the on-demand code execution system on a host device nearby to an end user, and thus execution of a task may require little more than provisioning the execution environment with code and executing that code. In other instances, these execution environments may not be executing, and thus execution of a task may also require initializing the environment (e.g., by launching a virtual machine instance, which may include loading an operating system, etc.). In general, maintaining an environment in an executing state can facilitate more rapid execution of a task, but also utilizes some amount of working computing resources of a host computing device, such as central processing unit (CPU) cycles and registers, random access memory (RAM), and the like. In contrast, maintaining an environment in a non-executing state (e.g., a shut down virtual machine instance), can utilize fewer or no working resources, and can instead utilize non-working resources (e.g., resources not required for currently executing processes) such as long term memory storage provided by a hard disk drive (HDD). However, maintaining an environment in a non-executing state may require that the environment first be initialized prior to executing a task within the environment, thus potentially delaying execution of the task.

To address this trade-off, embodiments of the on-demand code execution system can be configured to modify a memory state of an execution environment according to a next predicted execution of the task, such that the overall computing resource usage of the on-demand code execution system is reduced, and/or such that the on-demand code execution system can facilitate more executions of tasks within a given set of computing resources. Specifically, embodiments of the present disclosure enable the on-demand code execution system to utilize historical information regarding executions of tasks to predict, for a given task, when a next request to execute that task will occur. The on-demand code execution system may then place an execution environment for the task into a memory state based on that predicted next execution request. For example, where the next execution request is expected to occur shortly, the on-demand code execution system may maintain an environment for the task in an executing state, such as within RAM or other primary memory of a host device. Conversely, when the next execution request is expected to occur at a significantly later time, the on-demand code execution system may maintain the environment for the task in a non-executing state within a secondary memory, such as a hard disk drive. Because secondary memory is generally more readily available within a host device than primary memory, and because maintaining an environment in a non-executing state generally minimizes or negates load on processing resources of a host (e.g., a CPU), maintaining an environment within secondary memory can reduce computing resource usage of a host device and free those resources for use in executing other tasks, thus improving the efficiency of the on-demand code execution system overall.

In some instances, a host device of the on-demand code execution system may have access to multiple levels of secondary memory, in addition to a primary memory (e.g., RAM). For example, a host device may have access to very rapid non-volatile memory, such as 3D XPOINT™ memory, developed by Intel Corporation (which memory implements a specific technology architecture generally classified as a type of resistive random-access memory, or “ReRAM”), relatively less rapid flash storage (e.g., within a solid state disk (SSD) drive), relatively less rapid magnetic storage memory (e.g., within an HDD), and still relatively less rapid non-local storage (e.g., a network-attached storage, or “NAS,” which may be implemented by any of a variety of physical memory storage devices, including the preceding memory device types). In general, the speed of memory (e.g., in terms of bandwidth) can be expected to be inversely proportional to the amount of such memory available. Thus, 3D XPOINT memory is expected to be less available than flash storage, which is expected to be less available than magnetic storage, etc. As used herein, the term “lower tier” memory is generally intended (unless noted to the contrary) to refer to memory with lower speed but greater expected availability than a “higher tier” memory, which in turn is generally intended to refer to memory with higher speed and lower expected availability. As such, transitioning an execution environment to a lower tier of memory is generally expected to reduce the resource usage of the on-demand code execution system in maintaining that environment, while at the same time increasing the time required to initialize the environment such that a task may be executed within the environment. As will be discussed below, the on-demand code execution system can therefore be configured to transition an environment to a lowest tier memory possible while still maintaining the ability of the on-demand code execution system to rapidly execute a task within the environment based on a predicted next request to execute the task.

In one embodiment, a predicted next request to execute a task may be based on historical information regarding the task. Illustratively, if requests to execute a task have historically (e.g., over a past period of time, such as a day, week, month, year, etc.) occurred at a set frequency of once per minute, the on-demand code execution system may expect that a next request to execute the task will occur one minute after a prior request. This illustrative example may occur when an external system, such as a web service, has been configured to call to the on-demand code execution system for execution of a task at the set frequency. In some instances, the on-demand code execution system may be configured to calculate an expected range of time until a next request to execute a task based on historical information. For example, the on-demand code execution system may apply common statistical techniques to calculate a mean or median predicted time until a next execution, or a standard deviation of a distribution of times between requests to execute the code. As a further example, the on-demand code execution system may calculate an interquartile range of a historical distribution of times between calls to execute a task, and use this range (along with a known last call to execute the task) to predict when a next call to execute a task will occur. In other embodiments, the on-demand code execution system may utilize other inputs to predict a next expected request to execute a task. For example, where little or no history exists for a given task, the on-demand code execution system may instead utilize request histories of similar tasks to calculate the next expected request to execute the given task. Similar tasks may be identified, for example, based on length of source code for the task, functions called within the code, libraries utilized by the task, a type of environment (e.g., operating system or runtime environment) for the task, and the like. In some embodiments, the on-demand code execution system may have more specific knowledge of the next expected request to execute a task. For example, a creator of a task may explicitly ask that the on-demand code execution system execute the task at a set frequency, thus enabling the on-demand code execution system to determine with great accuracy when a next execution of the task will occur.

In one embodiment, the on-demand code execution system is configured, on detecting an idle execution environment for a task in a primary memory (e.g., after execution of the task within the environment), to transition the environment to a lowest tier memory which would still enable the environment to be transitioned back to an executing state prior to a time of a next expected request to execute the task. For example, where a next expected request to execute a task in an environment is one minute in the future, and transitioning an environment to a given tier of secondary memory and back into primary memory is expected to take 50 seconds of time (e.g., 25 seconds to halt execution and transition to secondary memory and the same amount of time to transition to primary memory and initialize the environment), the on-demand code execution system may be configured to conduct that transition, thus “freeing up” an amount of primary memory supporting execution of the environment for ten seconds of time. (The amount freed in practice would exceed this, as some amount of primary memory would be expected to become available even before transitioning of the environment out of primary memory completes, and some amount would be expected not to be utilized until transitioning of the environment back to primary memory completes.)

While the above example may reduce overall usage of the primary memory, this example may not be desirable for the on-demand code execution system overall, as transitioning execution environments to secondary memory may also incur costs in terms of resource usage. Illustratively, in the example above, transitioning an environment from primary memory to secondary memory over a period of 25 seconds may utilize both the secondary memory and bandwidth of a communication bus between the primary and secondary memory. These other resources (the secondary memory and communication bus) may also support other operations of the on-demand code execution system. Thus, while transitioning an environment to secondary memory may free up a portion of the primary memory for a ten second period, it may also incur costs in terms of use of the secondary memory and the communication bus.

To address concern, embodiments of the present disclosure may calculate an expected cost (e.g., in terms of usage of computing resources) associated with maintaining an execution environment in a primary memory, as well as expected costs for transitioning the environment to each potential secondary memory (e.g., in terms of both usage of the secondary memory and usage of resources, such as a communication bus, to move the environment to the secondary memory). Each cost may be based at least partly on a timing of a next expected request to execute a task within the environment. The on-demand code execution system may thereafter transition the environment to a memory tier with a lowest expected overall cost to the on-demand code execution system.

In one embodiment, the execution environments managed by the on-demand code execution system correspond to virtual machine instances. To transition such instances from an executing to non-executing state, the on-demand code execution system may utilize “snapshots” of such virtual machine instances. Snapshotting of virtual machines is a known technique, and thus will not be described in detail herein. However, in brief, snapshotting may generate a data file which stores a state of a virtual machine instance at a point in time, including state elements such as a content of CPU registers of the virtual machine instance, contents of RAM of the virtual machine instances, states of pages within RAM (e.g., as “dirty” or “clean”), and any other information required to return the virtual machine instances to its prior state at a later point in time. Thus, as will be described below, the on-demand code execution system may be configured to modify a memory state of a virtual machine instance from primary memory to secondary memory by snapshotting a current state of the instances, and placing that snapshot into secondary memory. The on-demand code execution system may further modify a memory state of a virtual machine instance from a secondary memory to primary memory by utilizing the snapshot to reinitialize the virtual machine image in an executing state.

In addition to transitioning a virtual machine instance between primary and secondary memory, embodiments of the present disclosure may further enable transitioning virtual machine instances between memories of different host computing devices. For example, an on-demand code execution system may be implemented by devices distributed across a number of geographic locations, such as multiple data centers, each of which includes a number of host devices. To facilitate rapid execution of code, the on-demand code execution system may be configured to attempt to execute a task requested by an end user within an environment on a host device nearby to the end user (e.g., geographically or in terms of network distance). Due to movement of users, changes in capacity, failure of machines, and the like, it may be desirable to move execution environments between host devices, such that when a request to execute a task is received from an end user, the execution environment is hosted on a host device nearby to an end user with sufficient capacity to execute the task. A number of techniques are known in the art to determine a host device to which an execution environment should be migrated, and these techniques will thus not be discussed in detail herein. After selecting a host device to which an execution environment should be migrated, embodiments described herein may be utilized to migrate the execution environment from a current host device to a destination host device. Illustratively, an execution environment may be migrated from an executing state on a first host device to a secondary memory of the second host device in accordance with embodiments of the present disclosure.

One potential issue raised by transitioning an execution environment into an inactive state, or transitioning an environment between host devices (e.g., migrating the environment) is a disruption of stateful network connections. Generally described, stateful network connections are communication sessions between two devices on a communication network, where each device maintains “state” information reflecting a state of the communication session. One example of a stateful network connection is a connection between two computing devices implementing the transport control protocol (TCP). In a TCP network connection, two devices maintain state information regarding a TCP session, and negotiate to establish the state of that session. The information maintained by each device generally includes a “5-tuple” (a set of five values that at least partly define the session) and a session number. The five values of the 5-tuple generally identify for the session a source internet protocol (IP) address and port number, a destination IP address and port number, and an identifier of TCP protocol. Each device further maintains a sequence number, identifying an ordering of communications in the session. If any of the information maintained at a device changes, the TCP connection can be broken. For example, if an execution environment establishes a TCP connection to an external device and is then migrated from one host device to another, the environment may obtain a new IP address based on its new location. This may alter the 5-tuple of the TCP session, resulting in the external device rejecting the TCP connection if the environment attempts to utilize it. Stateful connections can also be broken due to inactivity, such as a timeout occurring at either device. Thus, if an execution environment is rendered inactive for a sufficient period of time, an external device may consider a TCP connection with the environment to be closed. On reactivating the environment, the environment may not be aware of this change (since, from the point of view of the environment, the timeout might not have occurred), and may attempt to utilize the TCP connection, resulting in an error. While examples are provided herein with reference to TCP, similar issues may exist within any stateful or connection-oriented protocol, including protocols layered over otherwise stateless protocols (e.g., the User Datagram Protocol, or “UDP”). Thus, transitions in state or location of an execution environment can be detrimental to stateful network connections of the device.

Embodiments of the present disclosure address these problems by providing a mechanism for gracefully halting and recovering stateful network connections on an execution environment that is transitioned in state or location. Specifically, in accordance with embodiments of the present disclosure, a host device hosting an execution environment may, prior to transitioning the environment, notify external devices (e.g., devices to which the environment has a stateful network connection) that connection is ending. The host device may illustratively identify external devices based on a connection table maintained at the host device, and utilized to route network data from the execution environment to the external device. In one embodiment, the host device “masquerades” as the execution environment when notifying the external device that a connection is ending, such as by modifying data packets corresponding to the notification to appear to originate from the execution environment. Thus, the external device may recognize the connection has being terminated. The host device may further prevent communications from or to the execution environment, to prevent the environment from gaining knowledge that the external device believes the connection to be terminated (since such knowledge may cause the environment to attempt to reestablish the connection prematurely).

The host device may then transition the execution environment, either in state (e.g., to a secondary memory of the host device) or in location (e.g., to another host device). On reactivating the environment, the host device may notify the environment that each connection (believed by the environment to still be active) is terminating. In one embodiment, the host device “masquerades” as the external device when notifying the execution environment that a connection is ending, such as by modifying data packets corresponding to the notification to appear to originate from the external device. Thus, the environment may also recognize the connection as being terminated, synchronizing its knowledge of the connection with the external device.

Because the above interactions enable both the external device and the execution to have the same knowledge of the connection (as terminated), the environment may then reestablish the network connection to the external device, if necessary, while avoiding any extraneous communications based on non-synchronized knowledge (which would be expected to result in errors). Any stateful connections of the environment can therefore be gracefully reestablished after transitioning of the environment.

While alternative techniques may exist to gracefully maintain or reestablish stateful network connections on a transitioning execution environment, these techniques may not be preferable on an on-demand code execution system as disclosed herein. For example, during migration of a virtual machine, it may be possible to maintain network connections by utilizing virtual networking or tunneling technologies to redirect network traffic from a prior host device hosting an environment to a new host device. Illustratively, the prior host device may be configured to receive network packets directed to an environment from an external device, and to route the packets to the new host device for delivery to the environment. However, network redirection generally incurs additional latency and overhead in terms of compute resources used to facilitate such redirection. This additional latency and overhead can be particularly problematic in production environments intended to operate at low latencies, like an on-demand code execution system. Moreover, such tunneling or rerouting would generally be insufficient to handle reestablishment of stateful network connections in the instance that of an environment being transitioned to an inactive state for a relatively long duration of time (e.g., longer than a timeout value for the network connection). Thus, the embodiments disclosed herein for gracefully terminating and reestablishing network connections may be preferable to rerouting or tunneling techniques.

As will be appreciated by one of skill in the art in light of the present disclosure, the embodiments disclosed herein improves the ability of computing systems, such as on-demand code execution systems, to execute code in an efficient manner. Specifically, embodiments of the present disclosure increase the efficiency of computing resource usage of such systems by enabling execution environments to be transitioned to lower tier memory, while maintaining the ability of such systems to execute code rapidly in response to requests to do so. Further, embodiments of the present disclosure decrease the occurrence of errors in on-demand code execution systems, by enabling graceful recovery of stateful network connections in environments hosted by such systems. Moreover, the presently disclosed embodiments address technical problems inherent within computing systems; specifically, the limited nature of computing resources with which to execute code, the inefficiencies caused by maintaining unutilized environments in an executing state, and the difficulties of maintaining stateful network connections during transitions of execution environments. These technical problems are addressed by the various technical solutions described herein, including the selective transitioning of environments to lower tier memories based on a time until a next expected utilization of such an environment, and the graceful recovery of network connections for the environment by utilization of a host device to notify either or both an external device and the environment that each network connection has terminated. Thus, the present disclosure represents an improvement on existing data processing systems and computing systems in general.

The general execution of tasks on the on-demand code execution system will now be discussed. As described in detail herein, the on-demand code execution system may provide a network-accessible service enabling users to submit or designate computer-executable source code to be executed by virtual machine instances on the on-demand code execution system. Each set of code on the on-demand code execution system may define a “task,” and implement specific functionality corresponding to that task when executed on a virtual machine instance of the on-demand code execution system. Individual implementations of the task on the on-demand code execution system may be referred to as an “execution” of the task (or a “task execution”). The on-demand code execution system can further enable users to trigger execution of a task based on a variety of potential events, such as detecting new data at a network-based storage system, transmission of an application programming interface (“API”) call to the on-demand code execution system, or transmission of a specially formatted hypertext transport protocol (“HTTP”) packet to the on-demand code execution system. Thus, users may utilize the on-demand code execution system to execute any specified executable code “on-demand,” without requiring configuration or maintenance of the underlying hardware or infrastructure on which the code is executed. Further, the on-demand code execution system may be configured to execute tasks in a rapid manner (e.g., in under 100 milliseconds [ms]), thus enabling execution of tasks in “real-time” (e.g., with little or no perceptible delay to an end user). To enable this rapid execution, the on-demand code execution system can include one or more virtual machine instances that are “pre-warmed” or pre-initialized (e.g., booted into an operating system and executing a complete or substantially complete runtime environment) and configured to enable execution of user-defined code, such that the code may be rapidly executed in response to a request to execute the code, without delay caused by initializing the virtual machine instance. Thus, when an execution of a task is triggered, the code corresponding to that task can be executed within a pre-initialized virtual machine in a very short amount of time.

Specifically, to execute tasks, the on-demand code execution system described herein may maintain a pool of executing virtual machine instances that are ready for use as soon as a user request is received. Due to the pre-initialized nature of these virtual machines, delay (sometimes referred to as latency) associated with executing the user code (e.g., instance and language runtime startup time) can be significantly reduced, often to sub-100 millisecond levels. Illustratively, the on-demand code execution system may maintain a pool of virtual machine instances on one or more physical computing devices, where each virtual machine instance has one or more software components (e.g., operating systems, language runtimes, libraries, etc.) loaded thereon. When the on-demand code execution system receives a request to execute the program code of a user (a “task”), which specifies one or more computing constraints for executing the program code of the user, the on-demand code execution system may select a virtual machine instance for executing the program code of the user based on the one or more computing constraints specified by the request and cause the program code of the user to be executed on the selected virtual machine instance. The program codes can be executed in isolated containers that are created on the virtual machine instances, or may be executed within a virtual machine instance isolated from other virtual machine instances acting as environments for other tasks. Since the virtual machine instances in the pool have already been booted and loaded with particular operating systems and language runtimes by the time the requests are received, the delay associated with finding compute capacity that can handle the requests (e.g., by executing the user code in one or more containers created on the virtual machine instances) can be significantly reduced.

Because the number of different virtual machine instances that a host computing device may execute is limited by the computing resources of that host (and particularly by highly utilized resources such as CPU cycles and RAM), the number of virtual machine instances in a pool on the on-demand code execution system is similarly limited. Thus, in accordance with the embodiments of the present disclosure, the on-demand code execution system may generate execution environments for a large number of tasks (e.g., more environments than could be maintained as executing on the on-demand code execution system at a given point in time), and transition a subset (e.g., a majority) of those environments into lower tier memory storage, based on a next expected utilization of each environment. Thus, a primary memory of the on-demand code execution system can be expected to hold environments either being actively used or expected to be used in a very short amount of time. As environments within the primary memory become idle, the on-demand code execution system can transition the environments to secondary memory based on future expected use, and move into primary memory those environments which are next expected to be used. In this manner, the overall efficiency of primary memory within the on-demand code execution system is increased.

As used herein, the term “virtual machine instance” is intended to refer to an execution of software or other executable code that emulates hardware to provide an environment or platform on which software may execute (an “execution environment”). Virtual machine instances are generally executed by hardware devices, which may differ from the physical hardware emulated by the virtual machine instance. For example, a virtual machine may emulate a first type of processor and memory while being executed on a second type of processor and memory. Thus, virtual machines can be utilized to execute software intended for a first execution environment (e.g., a first operating system) on a physical device that is executing a second execution environment (e.g., a second operating system). In some instances, hardware emulated by a virtual machine instance may be the same or similar to hardware of an underlying device. For example, a device with a first type of processor may implement a plurality of virtual machine instances, each emulating an instance of that first type of processor. Thus, virtual machine instances can be used to divide a device into a number of logical sub-devices (each referred to as a “virtual machine instance”). While virtual machine instances can generally provide a level of abstraction away from the hardware of an underlying physical device, this abstraction is not required. For example, assume a device implements a plurality of virtual machine instances, each of which emulate hardware identical to that provided by the device. Under such a scenario, each virtual machine instance may allow a software application to execute code on the underlying hardware without translation, while maintaining a logical separation between software applications running on other virtual machine instances. This process, which is generally referred to as “native execution,” may be utilized to increase the speed or performance of virtual machine instances. Other techniques that allow direct utilization of underlying hardware, such as hardware pass-through techniques, may be used, as well.

While a virtual machine executing an operating system is described herein as one example of an execution environment, other execution environments are also possible. For example, tasks or other processes may be executed within a software “container,” which provides a runtime environment without itself providing virtualization of hardware. Containers may be implemented within virtual machines to provide additional security, or may be run outside of a virtual machine instance.

The foregoing aspects and many of the attendant advantages of this disclosure will become more readily appreciated as the same become better understood by reference to the following description, when taken in conjunction with the accompanying drawings.

FIG.1 is a block diagram of anillustrative operating environment100 in which an on-demandcode execution system110 may operate based on communication with user computing devices102,auxiliary services106, and network-based data storage services108. By way of illustration, various example user computing devices102 are shown in communication with the on-demandcode execution system110, including a desktop computer, laptop, and a mobile phone. In general, the user computing devices102 can be any computing device such as a desktop, laptop or tablet computer, personal computer, wearable computer, server, personal digital assistant (PDA), hybrid PDA/mobile phone, mobile phone, electronic book reader, set-top box, voice command device, camera, digital media player, and the like. The on-demandcode execution system110 may provide the user computing devices102 with one or more user interfaces, command-line interfaces (CLIs), application programing interfaces (APIs), and/or other programmatic interfaces for generating and uploading user-executable code (e.g., including metadata identifying dependency code objects for the uploaded code), invoking the user-provided code (e.g., submitting a request to execute the user codes on the on-demand code execution system110), scheduling event-based jobs or timed jobs, tracking the user-provided code, and/or viewing other logging or monitoring information related to their requests and/or user codes. Although one or more embodiments may be described herein as using a user interface, it should be appreciated that such embodiments may, additionally or alternatively, use any CLIs, APIs, or other programmatic interfaces.

Theillustrative environment100 further includes one or moreauxiliary services106, which can interact with the one-demandcode execution environment110 to implement desired functionality on behalf of a user.Auxiliary services106 can correspond to network-connected computing devices, such as servers, which generate data accessible to the one-demandcode execution environment110 or otherwise communicate to the one-demandcode execution environment110. For example, theauxiliary services106 can include web services (e.g., associated with the user computing devices102, with the on-demandcode execution system110, or with third parties), databases, really simple syndication (“RSS”) readers, social networking sites, or any other source of network-accessible service or data source. In some instances,auxiliary services106 may be associated with the on-demandcode execution system110, e.g., to provide billing or logging services to the on-demandcode execution system110. In some instances,auxiliary services106 actively transmit information, such as API calls or other task-triggering information, to the on-demandcode execution system110. In other instances,auxiliary services106 may be passive, such that data is made available for access by the on-demandcode execution system110. For example, components of the on-demandcode execution system110 may periodically poll such passive data sources, and trigger execution of tasks within the on-demandcode execution system110 based on the data provided. While depicted inFIG.1 as distinct from the user computing devices102 and the on-demandcode execution system110, in some embodiments, variousauxiliary services106 may be implemented by either the user computing devices102 or the on-demandcode execution system110.

Theillustrative environment100 further includes one or more network-based data storage services108, configured to enable the on-demandcode execution system110 to store and retrieve data from one or more persistent or substantially persistent data sources. Illustratively, the network-based data storage services108 may enable the on-demandcode execution system110 to store information corresponding to a task, such as code or metadata, to store additional code objects representing dependencies of tasks, to retrieve data to be processed during execution of a task, and to store information (e.g., results) regarding that execution. The network-based data storage services108 may represent, for example, a relational or non-relational database. In another example, the network-based data storage services108 may represent a network-attached storage (NAS), configured to provide access to data arranged as a file system. The network-based data storage services108 may further enable the on-demandcode execution system110 to query for and retrieve information regarding data stored within the on-demandcode execution system110, such as by querying for a number of relevant files or records, sizes of those files or records, file or record names, file or record creation times, etc. In some instances, the network-based data storage services108 may provide additional functionality, such as the ability to separate data into logical groups (e.g., groups associated with individual accounts, etc.). While shown as distinct from theauxiliary services106, the network-based data storage services108 may in some instances also represent a type ofauxiliary service106.

The user computing devices102,auxiliary services106, and network-based data storage services108 may communicate with the on-demandcode execution system110 via anetwork104, which may include any wired network, wireless network, or combination thereof. For example, thenetwork104 may be a personal area network, local area network, wide area network, over-the-air broadcast network (e.g., for radio or television), cable network, satellite network, cellular telephone network, or combination thereof. As a further example, thenetwork104 may be a publicly accessible network of linked networks, possibly operated by various distinct parties, such as the Internet. In some embodiments, thenetwork104 may be a private or semi-private network, such as a corporate or university intranet. Thenetwork104 may include one or more wireless networks, such as a Global System for Mobile Communications (GSM) network, a Code Division Multiple Access (CDMA) network, a Long Term Evolution (LTE) network, or any other type of wireless network. Thenetwork104 can use protocols and components for communicating via the Internet or any of the other aforementioned types of networks. For example, the protocols used by thenetwork104 may include Hypertext Transfer Protocol (HTTP), HTTP Secure (HTTPS), Message Queue Telemetry Transport (MQTT), Constrained Application Protocol (CoAP), and the like. Protocols and components for communicating via the Internet or any of the other aforementioned types of communication networks are well known to those skilled in the art and, thus, are not described in more detail herein.

The on-demandcode execution system110 is depicted inFIG.1 as operating in a distributed computing environment including several computer systems that are interconnected using one or more computer networks (not shown inFIG.1). The on-demandcode execution system110 could also operate within a computing environment having a fewer or greater number of devices than are illustrated inFIG.1. Thus, the depiction of the on-demandcode execution system110 inFIG.1 should be taken as illustrative and not limiting to the present disclosure. For example, the on-demandcode execution system110 or various constituents thereof could implement various Web services components, hosted or “cloud” computing environments, and/or peer to peer network configurations to implement at least a portion of the processes described herein.

Further, the on-demandcode execution system110 may be implemented directly in hardware or software executed by hardware devices and may, for instance, include one or more physical or virtual servers implemented on physical computer hardware configured to execute computer executable instructions for performing various features that will be described herein. The one or more servers may be geographically dispersed or geographically co-located, for instance, in one or more data centers. In some instances, the one or more servers may operate as part of a system of rapidly provisioned and released computing resources, often referred to as a “cloud computing environment.”

In the example ofFIG.1, the on-demandcode execution system110 is illustrated as connected to thenetwork104. In some embodiments, any of the components within the on-demandcode execution system110 can communicate with other components of the on-demandcode execution system110 via thenetwork104. In other embodiments, not all components of the on-demandcode execution system110 are capable of communicating with other components of thevirtual environment100. In one example, only the frontend120 (which may in some instances represent multiple frontends120) may be connected to thenetwork104, and other components of the on-demandcode execution system110 may communicate with other components of theenvironment100 via thefrontends120.

InFIG.1, users, by way of user computing devices102, may interact with the on-demandcode execution system110 to provide executable code, and establish rules or logic defining when and how such code should be executed on the on-demandcode execution system110, thus establishing a “task.” For example, a user may wish to run a piece of code in connection with a web or mobile application that the user has developed. One way of running the code would be to acquire virtual machine instances from service providers who provide infrastructure as a service, configure the virtual machine instances to suit the user's needs, and use the configured virtual machine instances to run the code. In order to avoid the complexity of this process, the user may alternatively provide the code to the on-demandcode execution system110, and request that the on-demandcode execution system110 execute the code. The on-demandcode execution system110 can handle the acquisition and configuration of compute capacity (e.g., containers, instances, etc., which are described in greater detail below) based on the code execution request, and execute the code using the compute capacity. The on-demandcode execution system110 may automatically scale up and down based on the volume, thereby relieving the user from the burden of having to worry about over-utilization (e.g., acquiring too little computing resources and suffering performance issues) or under-utilization (e.g., acquiring more computing resources than necessary to run the codes, and thus overpaying). In accordance with embodiments of the present disclosure, and as described in more detail below, the on-demandcode execution system110 may configure the virtual machine instances with customized operating systems to execute the user's code more efficiency and reduce utilization of computing resources.

To enable interaction with the on-demandcode execution system110, thesystem110 includes one ormore frontends120, which enable interaction with the on-demandcode execution system110. In an illustrative embodiment, thefrontends120 serve as a “front door” to the other services provided by the on-demandcode execution system110, enabling users (via user computing devices102) to provide, request execution of, and view results of computer executable code. Thefrontends120 include a variety of components to enable interaction between the on-demandcode execution system110 and other computing devices. For example, each frontend120 may include a request interface providing user computing devices102 with the ability to upload or otherwise communication user-specified code to the on-demandcode execution system110 and to thereafter request execution of that code. In one embodiment, the request interface communicates with external computing devices (e.g., user computing devices102,auxiliary services106, etc.) via a graphical user interface (GUI), CLI, or API. Thefrontends120 process the requests and makes sure that the requests are properly authorized. For example, thefrontends120 may determine whether the user associated with the request is authorized to access the user code specified in the request.

References to user code as used herein may refer to any program code (e.g., a program, routine, subroutine, thread, etc.) written in a specific program language. In the present disclosure, the terms “code,” “user code,” and “program code,” may be used interchangeably. Such user code may be executed to achieve a specific function, for example, in connection with a particular web application or mobile application developed by the user. As noted above, individual collections of user code (e.g., to achieve a specific function) are referred to herein as “tasks,” while specific executions of that code (including, e.g., compiling code, interpreting code, or otherwise making the code executable) are referred to as “task executions” or simply “executions.” Tasks may be written, by way of non-limiting example, in JavaScript (e.g., node.js), Java, Python, and/or Ruby (and/or another programming language). Tasks may be “triggered” for execution on the on-demandcode execution system110 in a variety of manners. In one embodiment, a user or other computing device may transmit a request to execute a task may, which can generally be referred to as “call” to execute of the task. Such calls may include the user code (or the location thereof) to be executed and one or more arguments to be used for executing the user code. For example, a call may provide the user code of a task along with the request to execute the task. In another example, a call may identify a previously uploaded task by its name or an identifier. In yet another example, code corresponding to a task may be included in a call for the task, as well as being uploaded in a separate location (e.g., storage of anauxiliary service106 or a storage system internal to the on-demand code execution system110) prior to the request being received by the on-demandcode execution system110. As noted above, the code for a task may reference additional code objects maintained at the on-demandcode execution system110 by use of identifiers of those code objects, such that the code objects are combined with the code of a task in an execution environment prior to execution of the task. The on-demandcode execution system110 may vary its execution strategy for a task based on where the code of the task is available at the time a call for the task is processed. A request interface of thefrontend120 may receive calls to execute tasks as Hypertext Transfer Protocol Secure (HTTPS) requests from a user. Also, any information (e.g., headers and parameters) included in the HTTPS request may also be processed and utilized when executing a task. As discussed above, any other protocols, including, for example, HTTP, MQTT, and CoAP, may be used to transfer the message containing a task call to the request interface122.

A call to execute a task (which may also be referred to as a request to execute the task) may specify one or more third-party libraries (including native libraries) to be used along with the user code corresponding to the task. In one embodiment, the call may provide to the on-demand code execution system110 a file containing the user code and any libraries (and/or identifications of storage locations thereof) corresponding to the task requested for execution. In some embodiments, the call includes metadata that indicates the program code of the task to be executed, the language in which the program code is written, the user associated with the call, and/or the computing resources (e.g., memory, etc.) to be reserved for executing the program code. For example, the program code of a task may be provided with the call, previously uploaded by the user, provided by the on-demand code execution system110 (e.g., standard routines), and/or provided by third parties. Illustratively, code not included within a call or previously uploaded by the user may be referenced within metadata of the task by use of a URI associated with the code. In some embodiments, such resource-level constraints (e.g., how much memory is to be allocated for executing a particular user code) are specified for the particular task, and may not vary over each execution of the task. In such cases, the on-demandcode execution system110 may have access to such resource-level constraints before each individual call is received, and the individual call may not specify such resource-level constraints. In some embodiments, the call may specify other constraints such as permission data that indicates what kind of permissions or authorities that the call invokes to execute the task. Such permission data may be used by the on-demandcode execution system110 to access private resources (e.g., on a private network). In some embodiments, individual code objects may also be associated with permissions or authorizations. For example, a third party may submit a code object and designate the object as readable by only a subset of users. The on-demandcode execution system110 may include functionality to enforce these permissions or authorizations with respect to code objects.

In some embodiments, a call may specify the behavior that should be adopted for handling the call. In such embodiments, the call may include an indicator for enabling one or more execution modes in which to execute the task referenced in the call. For example, the call may include a flag or a header for indicating whether the task should be executed in a debug mode in which the debugging and/or logging output that may be generated in connection with the execution of the task is provided back to the user (e.g., via a console user interface). In such an example, the on-demandcode execution system110 may inspect the call and look for the flag or the header, and if it is present, the on-demandcode execution system110 may modify the behavior (e.g., logging facilities) of the container in which the task is executed, and cause the output data to be provided back to the user. In some embodiments, the behavior/mode indicators are added to the call by the user interface provided to the user by the on-demandcode execution system110. Other features such as source code profiling, remote debugging, etc. may also be enabled or disabled based on the indication provided in a call.

To manage requests for code execution, thefrontend120 can include an execution queue (not shown inFIG.1), which can maintain a record of requested task executions. Illustratively, the number of simultaneous task executions by the on-demandcode execution system110 is limited, and as such, new task executions initiated at the on-demand code execution system110 (e.g., via an API call, via a call from an executed or executing task, etc.) may be placed on the execution queue124 and processed, e.g., in a first-in-first-out order. In some embodiments, the on-demandcode execution system110 may include multiple execution queues, such as individual execution queues for each user account. For example, users of the on-demandcode execution system110 may desire to limit the rate of task executions on the on-demand code execution system110 (e.g., for cost reasons). Thus, the on-demandcode execution system110 may utilize an account-specific execution queue to throttle the rate of simultaneous task executions by a specific user account. In some instances, the on-demandcode execution system110 may prioritize task executions, such that task executions of specific accounts or of specified priorities bypass or are prioritized within the execution queue. In other instances, the on-demandcode execution system110 may execute tasks immediately or substantially immediately after receiving a call for that task, and thus, the execution queue may be omitted.

As noted above, tasks may be triggered for execution at the on-demandcode execution system110 based on explicit calls from user computing devices102 (e.g., as received at the request interface). Alternatively or additionally, tasks may be triggered for execution at the on-demandcode execution system110 based on data retrieved from one or moreauxiliary services106 or network-based data storage services108. To facilitate interaction withauxiliary services106, thefrontend120 can include a polling interface (not shown inFIG.1), which operates to pollauxiliary services106 or data storage services108 for data. Illustratively, the polling interface may periodically transmit a request to one or more user-specifiedauxiliary services106 or data storage services108 to retrieve any newly available data (e.g., social network “posts,” news articles, files, records, etc.), and to determine whether that data corresponds to a user-established criteria triggering execution a task on the on-demandcode execution system110. Illustratively, criteria for execution of a task may include, but is not limited to, whether new data is available at theauxiliary services106 or data storage services108, the type or content of the data, or timing information corresponding to the data. In some instances, theauxiliary services106 or data storage services108 may function to notify thefrontend120 of the availability of new data, and thus the polling service may be unnecessary with respect to such services.

In addition to tasks executed based on explicit user calls and data fromauxiliary services106, the on-demandcode execution system110 may in some instances operate to trigger execution of tasks independently. For example, the on-demandcode execution system110 may operate (based on instructions from a user) to trigger execution of a task at each of a number of specified time intervals (e.g., every 10 minutes).

Thefrontend120 can further include an output interface (not shown inFIG.1) configured to output information regarding the execution of tasks on the on-demandcode execution system110. Illustratively, the output interface may transmit data regarding task executions (e.g., results of a task, errors related to the task execution, or details of the task execution, such as total time required to complete the execution, total data processed via the execution, etc.) to the user computing devices102 or toauxiliary services106, which may include, for example, billing or logging services. The output interface may further enable transmission of data, such as service calls, toauxiliary services106. For example, the output interface may be utilized during execution of a task to transmit an API request to an external service106 (e.g., to store data generated during execution of the task).

In some embodiments, the on-demandcode execution system110 may includemultiple frontends120. In such embodiments, a load balancer (not shown inFIG.1) may be provided to distribute the incoming calls to themultiple frontends120, for example, in a round-robin fashion. In some embodiments, the manner in which the load balancer distributes incoming calls to themultiple frontends120 may be based on the location or state of other components of the on-demandcode execution system110. For example, a load balancer may distribute calls to a geographicallynearby frontend120, or to a frontend with capacity to service the call. In instances where eachfrontend120 corresponds to an individual instance of another component of the on-demand code execution system, such as theactive pool148 described below, the load balancer may distribute calls according to the capacities or loads on those other components. Calls may in some instances be distributed betweenfrontends120 deterministically, such that a given call to execute a task will always (or almost always) be routed to thesame frontend120. This may, for example, assist in maintaining an accurate execution record for a task, to ensure that the task executes only a desired number of times. For example, calls may be distributed to load balance betweenfrontend120. Other distribution techniques, such as anycast routing, will be apparent to those of skill in the art.

The on-demand code execution system further includes one ormore worker managers140 that manage the execution environments, such as virtual machine instances150 (shown asVM instance150A and150B, generally referred to as a “VM”), used for servicing incoming calls to execute tasks, and that manage the memory states of execution environments. While the following will be described with reference tovirtual machine instances150 as examples of such environments, embodiments of the present disclosure may utilize other environments, such as software containers. In the example illustrated inFIG.1, eachworker manager140 manages anactive pool148, which is a group (sometimes referred to as a pool) ofvirtual machine instances150 executing on one or more physical host computing devices that are initialized to execute a given task (e.g., by having the code of the task and any dependency data objects loaded into the instance). Theactive pool148 illustratively is implemented using primary memory (e.g., RAM) of host devices implementing or under control of theworker manager140.

Although thevirtual machine instances150 are described here as being assigned to a particular task, in some embodiments, the instances may be assigned to a group of tasks, such that the instance is tied to the group of tasks and any tasks of the group can be executed within the instance. For example, the tasks in the same group may belong to the same security group (e.g., based on their security credentials) such that executing one task in a container on aparticular instance150 after another task has been executed in another container on the same instance does not pose security risks. As another example, the tasks of the group may share common dependencies, such that an environment used to execute one task of the group can be rapidly modified to support execution of another task within the group.

Once a triggering event to execute a task has been successfully processed by afrontend120, thefrontend120 passes a request to aworker manager140 to execute the task. In one embodiment, each frontend120 may be associated with a corresponding worker manager140 (e.g., aworker manager140 co-located or geographically nearby to the frontend120) and thus, thefrontend120 may pass most or all requests to thatworker manager140. In another embodiment, afrontend120 may include a location selector configured to determine aworker manager140 to which to pass the execution request. In one embodiment, the location selector may determine theworker manager140 to receive a call based on hashing the call, and distributing the call to aworker manager140 selected based on the hashed value (e.g., via a hash ring). Various other mechanisms for distributing calls betweenworker managers140 will be apparent to one of skill in the art.

Thereafter, theworker manager140 may modify a virtual machine instance150 (if necessary) and execute the code of the task within theinstance150. As shown inFIG.1,respective instances150 may have operating systems (OS)152 (shown asOS152A and152B), language runtimes154 (shown asruntime154A and154B), and user code156 (shown as user code156A and156B). The OS152, runtime154, and user code156 may collectively enable execution of the user code to implement the task. In some instances, eachVM150 may be associated with additional information, such as state information, maintained across individual executions of a task. For example, when initially created, aVM150 may initialize the OS152, and each time the user code156 is executed in theVM150, a state of theVM150 may change. State of aVM150 may be maintained, for example, within registers of a virtual CPU of theVM150, within RAM of theVM150, within a virtual disk drive of theVM150, or the like.

In accordance with embodiments of the present disclosure, theworker manager140 further includes amanagement unit142, configured to manage a state of theVMs150. Specifically, themanagement unit142 may be configured to detect anidle VM150 within theactive pool148 and to determine, based on an expected next use of thatVM150, whether to transition theVM150 to asecondary memory144. Thesecondary memories144 can correspond to one or more lower tier memories, which are less rapid than primary memory, but which are illustratively greater in capacity. Thesecondary memories144 can correspond, for example, to 3D XPOINT, flash memory, magnetic storage, or network-attached storage. Specifically, themanagement unit142 can be configured to calculate a next expected use of a VM150 (e.g., when a next request to execute a task within theVM150 will be received), and to calculate both (i) a cost of maintaining theVM150 within theactive pool148 until that next expected use and (ii) a cost of transitioning theVM150 to one of thesecondary memories144 and back into theactive pool148 so that it is available at the time of the next expected use. If the cost of transitioning theVM150 into asecondary memory144 is less than a cost of maintaining theVM150 within theactive pool148, themanagement unit142 can transition theVM150 into thesecondary memory144. Themanagement unit142 can later transition theVM150 back into theactive pool148 as required to service a next request to execute a task in theVM150. For example, themanagement unit142 may time a transition of theVM150 from thesecondary memory144 to theactive pool148 such that the transition completes at or just prior to an expected time of the request. In instances where a request is received before that expected time, themanagement unit142 may transition theVM150 earlier, such as immediately (e.g., on receive of the request) or as space is available within theactive pool148.

To assist in transitioning ofVMs150 betweenactive pool148 andsecondary memory144, theworker manager140 ofFIG.1 further includes amemory utilization monitor146, configured to monitor memory utilization ofvarious VMs150 on the worker manager140 (e.g., on one or more host devices implementing the worker manager140). In one embodiment, the memory utilization monitor146 monitors the memory utilization of aVM150 within a primary memory of theworker manager140. In another embodiment, the memory utilization monitor146 monitors an amount of difference (or “delta”) between a memory of aVM150 in primary memory of theworker manager140 and a representation of thatVM150 withinsecondary memory144. For example, in some embodiments, theworker manager140 may maintain in secondary memory144 a snapshot (or other data representation) of aVM150, even when thatVM150 is executing within theactive pool148. Illustratively, the snapshot may have been previously utilized to generate theVM150 within theactive pool148. The memory state of theVM150 within theactive pool148 and the snapshot of theVM150 in thesecondary memory144 may include a substantial amount of information. For example, aVM150 may be loaded with a runtime154 which, on initialization, utilizes a given set of memory space (e.g., 100 megabytes). When executing user code156 in the runtime154, some subset of that memory space may be modified based on execution of the user code (e.g., 10 megabytes). This memory utilized during execution of a task can represent the “working set” of memory of the task. Thus, if a snapshot of theVM150 is taken just after initialization and prior to execution of user code156, the snapshot and a state of theVM150 after execution of a task would be expected to overlap by 90 megabytes (the initial utilization minus the working set). Due to this overlap, a new snapshot of theVM150 after execution of a task may be created in thesecondary memory144 based on transferring out of the working set of theVM150 within theactive pool148, without requiring that all memory of theVM150 be transferred from theactive pool148. The remaining memory of theVM150 may be identified by reference to the preexisting snapshot of theVM150. Such interdependent snapshots are generally referred to as “cascading snapshots.” Thus, the memory utilization monitor146 in one configuration of thesystem110 tracks the “delta” of data needed to transition aVM150 from theactive pool148 to asecondary memory144.

As noted above, themanagement unit142 can be configured to determine when to transition aVM150 tosecondary memory144 based on a next expected utilization of theVM150 to execute a task. To facilitate determination of the next expected utilization, thesystem110 further includes a call history data store164, which stores information regarding a history of calls to thesystem110 of tasks. Illustratively, each time thefrontend120 receives a call to execute a task, thefrontend120 may update a record within the call history data store164 (e.g., within a database) to reflect that call. Themanagement unit142 may utilize the call history of a task to predict a next execution of that task or similar tasks. In some instances, themanagement unit142 may generate statistical information regarding the call history of a task, such as a median or mean duration between calls, a standard deviation of that duration, an interquartile range of that duration, or the like. Such statistical information may also be stored within the call history data store164. The call history data store164 may correspond to any persistent or substantially persistent data storage device, including (but not limited to) hard disk drives, solid state disk drives, network attached storage, etc., or any combination thereof.

In addition, themanagement unit142 may further be configured to facilitate graceful termination of stateful network connections ofVM instances150 in connection with transitioning of theVM instances150 in memory state or location. Specifically, as will be described below, themanagement unit142 may be configured to determine a set of stateful network connections of aVM instance150, such as by inspecting a connection table or data set maintained by a host device hosting theVM instance150. Themanagement unit142 may then transmit, to each “end” of each connection (e.g., theinstance150 and a device external to the instance, such as anotherinstance150, anauxiliary service106, etc.) a termination signal indicating that the connection has terminated. Subsequent to reactivation of theinstance150, a task executing on theinstance150 may then function to reestablish each required network connection, based on shared knowledge between theinstance150 and each external device that the respective connections have been terminated.

In one embodiment, themanagement unit142 notifies external devices of termination of each connection prior to or during deactivation of aVM instance150 on a host device initially hosting the instance. Notifying external devices of termination during or prior to deactivation may be beneficial, for example, in allowing external devices to close the network connections, rather than consider such connections “open” until a timeout event occurs. Moreover, when aVM instance150 is transitioning locations (e.g., migrating from a first to a second host device), notifying external devices from the first host device may beneficially increase the likelihood that a termination signal is accepted by the external device as authentic, since the first host is the same physical computing device that has previously transmitted data to the external device on behalf of the VM instance150 (and, for example, has been assigned the IP address previously used in communications with the external device).

Conversely, themanagement unit142 may be configured to notify theVM instance150 of termination of each connection during re-activation of the VM instance150 (e.g., as a final stage of migration to a second host, during re-inclusion in theactive pool148, etc.). Illustratively, theVM instance150 may be halted at a point at which it is executing code of a task, which code is configured on detection of a connection termination to attempt to reestablish the connection. Thus, if a termination signal were transmitted to theVM instance150 prior to halting, a “race condition” might occur, where theVM instance150 attempts to reestablish a network connection at the same time that themanagement unit142 is attempting to transition theVM instance150. To avoid this, themanagement unit142 may halt theVM instance150 without notifying theinstance150 of any network connection termination. When theVM instance150 is halted without receiving a termination of the connection, the code may assume that the connection is active. Thus, on reactivation of theVM instance150, themanagement unit142 may notify theVM instance150 that each connection has been terminated, thus causing theVM instance150 to reestablish those connections as necessary.

While illustrative timings for transmissions of termination signals are discussed above, these timings may be modified in embodiments of the present disclosure. For example, amanagement unit142 may transmit termination signals to both external devices and aVM instance150 prior to removing theinstance150 from anactive pool148. In this embodiment (or additionally in the other embodiments disclosed herein), themanagement unit142 may be configured to block communications between theVM instance150 and external devices, to prevent theinstance150 or external device from reestablishing a network connection prior to transitioning theinstance150. Additionally or alternatively, themanagement unit142 may transmit termination signals to both external devices and aVM instance150 after transitioning theinstance150 back into anactive pool148.

In some embodiments, themanagement unit142 may be configured to transmit termination signals to each end of a stateful network connection of aVM instance150 at least partly by masquerading as the opposing end of the connection. Specifically, themanagement unit142 may execute to control operation of a hypervisor, “DOM 0,” or lower-level operating system of a host device hosting aVM instance150, which may necessarily (by virtue of hosting the instance150) act as a “middle man” between theVM instance150 and each external device. Due to this position, the host device may generate packets for transmission to an external device that are similar or identical to those that aninstance150 would transmit to the external device. Similarly, the host device may generate packets for transmission to aninstance150 that are similar or identical to those that an external device would transmit to theinstance150. Themanagement unit142 may thus control operation of a host device to “masquerade” as one party to a network connection with speaking with the other party, increasing the likelihood that the spoken-to party accepts a transmission as legitimate.

While some functionalities are generally described herein with reference to an individual component of the on-demandcode execution system110, other components or a combination of components may additionally or alternatively implement such functionalities. For example, while eachworker manager140 is depicted inFIG.1 as including amanagement unit142, in some instances, one or more centralizedstate management units142 may be provided which provide access to the above-described functionalities tomultiple worker managers140. In instances where theworker manager140 is implemented across multiple host devices, each host device may include amanagement unit142, or multiple devices may share access to a common management unit142 (e.g., executed on a physically nearby device, such as a device on the same “rack” within a data center). In some instances, the call history data store164 may be implemented locally within aworker manager140. For example, where eachworker manager140 is configured to handle a specific subset of tasks on thesystem110, eachworker manager140 may maintain a call history of that specific subset of tasks. Thus, the specific configuration of elements withinFIG.1 is intended to be illustrative.

FIG.2 depicts a general architecture of a computing system implementing aworker manager140 ofFIG.1. The general architecture of theworker manager140 depicted inFIG.2 includes an arrangement of computer hardware and software that may be used to implement aspects of the present disclosure. The hardware may be implemented on physical electronic devices, as discussed in greater detail below. Theworker manager140 may include many more (or fewer) elements than those shown inFIG.2. It is not necessary, however, that all of these generally conventional elements be shown in order to provide an enabling disclosure. Additionally, the general architecture illustrated inFIG.2 may be used to implement one or more of the other components illustrated inFIG.1.

As illustrated, theworker manager140 includes aprocessing unit290, anetwork interface292, a computerreadable medium drive294, and an input/output device interface296, all of which may communicate with one another by way of a communication bus. Thenetwork interface292 may provide connectivity to one or more networks or computing systems. Theprocessing unit290 may thus receive information and instructions from other computing systems or services via thenetwork104. Theprocessing unit290 may also communicate to and fromprimary memory280 and/orsecondary memory298 and further provide output information for an optional display (not shown) via the input/output device interface296. The input/output device interface296 may also accept input from an optional input device (not shown).

Theprimary memory280 and/orsecondary memory298 may contain computer program instructions (grouped as units in some embodiments) that theprocessing unit290 executes in order to implement one or more aspects of the present disclosure. These program instructions are shown inFIG.2 as included within theprimary memory280, but may additionally or alternatively be stored withinsecondary memory298. Theprimary memory280 andsecondary memory298 correspond to one or more tiers of memory devices, including (but not limited to) RAM, 3D XPOINT memory, flash memory, magnetic storage, and the like. Theprimary memory280 is assumed for the purposes of description to represent a main working memory of theworker manager140, with a higher speed but lower total capacity thansecondary memory298. As noted above, thesecondary memory298 may include multiple tiers of memory, each lower representing a progressively lower speed but potentially higher capacity than a prior tier.

Theprimary memory280 illustratively includes theactive pool148, which may store information regarding virtual machine instances that are actively executing on theworker manager140. While shown as part ofprimary memory280, use of the term “active pool” may in some instances also refer to a logical construct including theVMs150 executing on aworker manager140. TheseVMs150 within theactive pool148 may additionally utilize other resources of theworker manager140, such as theprocessing unit290, thenetwork interface292, etc. Thus, inclusion of the “active pool148” withinprimary memory280 is intended to visually represent a portion of theprimary memory280 utilized byVMs150 within theactive pool148, and not as an indication thatsuch VMs150 solely utilize theprimary memory280.

Theprimary memory280 may further store anoperating system284 that provides computer program instructions for use by theprocessing unit290 in the general administration and operation of the sidecar configuration system160. Thememory280 may further include computer program instructions and other information for implementing aspects of the present disclosure. For example, in one embodiment, thememory280 includes auser interface unit282 that generates user interfaces (and/or instructions therefor) for display upon a computing device, e.g., via a navigation and/or browsing interface such as a browser or application installed on the computing device. In addition, thememory280 may include and/or communicate with one or more data repositories (not shown), for example, to access user program codes and/or libraries.

In addition to and/or in combination with theuser interface unit282, thememory280 may include a virtual machine configuration unit162,management unit142, andmemory utilization monitor146. In one embodiment, the virtual machine configuration unit162,management unit142, and memory utilization monitor146 individually or collectively implement various aspects of the present disclosure, e.g., generating virtual machine instances in which to execute code in response to requests for such execution, monitoring memory usage of such machines within theprimary memory280, and selectively transitioningVMs150 out ofprimary memory280 and intosecondary memory298 based on a next expected use of theVM150 to execute a task within thesystem110.

Theworker manager140 ofFIG.2 is one illustrative configuration of such a device, of which others are possible. For example, while shown as a single device, aworker manager140 may in some embodiments be implemented as multiple physical host devices. Illustratively, a first device of such aworker manager140 may implement “control plane” functions, such as receiving requests to execute tasks, instructing when to transitionVMs150 between memory states, and the like, while a second device may house theactive pool148 and implement “data plane” operations, such as executing code in response to instructions by the first device.

In some embodiments, theworker manager140 may further include components other than those illustrated inFIG.2. For example, theprimary memory280 may further include a container manager for managing creation, preparation, and configuration of containers within virtual machine instances. Further variations on functionalities of aworker manager140, any of which may be implanted by theworker manager140 of the present disclosure, are described in more detail in U.S. Pat. No. 9,323,556, entitled “PROGRAMMATIC EVENT DETECTION AND MESSAGE GENERATION FOR REQUESTS TO EXECUTE PROGRAM CODE,” and filed Sep. 30, 2014 (the “'556 Patent”), the entirety of which is hereby incorporated by reference.

With reference toFIG.3, illustrative interactions are depicted for determining whether to transition the memory state of aVM150 based on a next expected use of theVM150 to execute a task. The interactions ofFIG.3 begin at (1), where themanagement unit142 detects anidle VM150. In the present description, anidle VM150 is intended to refer to aVM150 that is not actively servicing a request to execute a task, such as by executing code of that task. For ease of description, it will be assumed with reference toFIG.3 that eachVM150 is dedicated to execution of a single task. However, the interactions ofFIG.3 may be modified to includeVMs150 made available to execute multiple tasks (e.g., of the same user, utilizing the same libraries, etc.). Illustratively, themanagement unit142 may detect anidle VM150 by receiving an indication that theVM150 has completed execution of a task, and determining that no outstanding requests for execution of that task exist on theworker manager140.

Thereafter, at (2) and (3) (which interactions are described in sequence but may be implemented concurrently, in parallel, in the reverse order, etc.), themanagement unit142 retrieves task call history information from the call history data store164, and VM memory usage information from thememory utilization monitor146, respectively. As noted above, the task call history information may include a record of requests to execute a task (or one or more similar tasks) and/or statistical information related to such a record. The VM memory usage information can include a current amount of primary memory utilized by aVM150, and/or a delta indicating how much information must be transitioned out of primary memory to transition theVM150's state.

At (5), themanagement unit142 determines a timing of a next expected call to execute a task within theVM150, based on the call history information. In instances where the call history information shows variations in the durations between calls, themanagement unit142 can use a statistical measurement to calculate an expected time until a next call to execute the task. In one embodiment, the expected time is calculated based on a probability that, if a next call conforms to the distribution of past calls, that the next call will occur no earlier than a given point in time. Illustratively, a statistical analysis of the call history information may indicate that there is a 99% chance according to the historical distribution of calls that a next call occurs no earlier than 10 seconds from the current point in time, a 90% chance that the next call occurs no earlier than 30 seconds from the current point in time, a 50% chance that the next call occurs no earlier than 60 seconds from the current point in time, etc. As such, theworker manager140 may be configured to utilize such a probability threshold to establish an expected timing of a next call. For example, theworker manager140 may be configured to utilize an expected timing such that there is a predicted n % chance that the next call occurs no earlier than the expected timing. In some instances, the specific percentage may vary based on a user associated with the task executed within theVM150 at issue. For example, specific users may require very low latency completion of requests, and thus may establish a high percentage value for calculating a next expected call. This can generally reduce the expected timing of a next call, and inhibit transferring of aVM150 out of theactive pool144. Conversely, other users may not require low latency completion of requests, and may thus establish a low percentage value for calculating a next expected call, leading to more aggressive transferring of aVM150 tosecondary memory144. In either instance, should a request to execute a task come before the next expected utilization, theworker manager140 can transition theVM150 out ofsecondary memory144 at the time of the request. Thus, such requests may still be serviced, but may incur additional latency due to the need to transition theVM150 out ofsecondary memory144 at the time of the request.

At (4), themanagement unit142 calculates expected costs for maintaining theVM150 within theactive pool148, as well as for transitioning theVM150 into asecondary memory144. Generally described, the cost of maintaining aVM150 within the active pool can be calculated as a sum of the resources of theworker manager140 to maintain theVM150 within theactive pool148 for the period of time until a next expected call to execute a task within theVM150. For example, where an executingVM150 within the active pool utilizes CPU cycles and RAM, the cost of maintaining theVM150 within theactive pool148 can represent a combination of total CPU cycles expected to be utilized until the next expected call and a metric reflecting RAM usage over time (e.g., “megabyte-seconds”). Theworker manager140 can be configured to weight each metric according to a relative scarcity of the corresponding resource, and combine the metrics via a weighted linear equation. For example, where aworker manager140 has excess CPU cycles but limited RAM, the metric of RAM usage over time may be weighted more highly than CPU cycles. A combination of these two metrics (potentially along with other metrics reflecting other resources used by an executing VM150) can represent the cost of maintaining theVM150 within the active pool.

Similarly, the cost of transitioning aVM150 tosecondary memory144 can reflected a weighted combination of the resources needed to transition theVM150 to thesecondary memory144 from the active pool, and later to transition theVM150 back from thesecondary memory144 to the active pool to service a next expected request. Such resources can include (but are not limited to) the amount of primary memory used over time to transition theVM150 in or out of the primary pool, the amount of a communication bus to transition the VM150 (e.g., the bandwidth over time, which may in some instances be represented as a total amount of data transferred over the bus), and the amount ofsecondary memory144 used over time to store theVM150 within thesecondary memory144. Each resources may be weighted according to the relative scarcity of the resource within theworker manager140. The total expected use of each resource may further be based on the expected timing of each phase of a transition of aVM150 intosecondary memory144, including a transition-out phase (e.g., using resources of a primary memory, asecondary memory144, and a communication bus between the two to transition aVM150 from theactive pool148 to the secondary memory144), an idle phase (e.g., using resources of thesecondary memory144 to maintain theVM150 in the secondary memory144), and a transition-in phase (e.g., using resources of a primary memory, asecondary memory144, and a communication bus between the two to transition aVM150 from thesecondary memory144 to the active pool148). Theworker manager140 can determine an expected timing of transition-in or transition-out phases based on a theoretical or observed timing of transfer of data between primary and secondary memories, as well as the amount of data needed to be transferred to transfer theVM150. For example, where 10 megabytes of data is required to be moved between a primary memory and asecondary memory144 to transfer a VM, and a communication bus between those two memories has an expected speed of 1 gigabyte per second (GB/s) (which may be determined either based on specification of the bus or based on historical observation of bus speed with respect to VM data), the expected transfer-out time may be one one-hundredth of a second (1 GB/s divided by 10 MB).

As noted above, the total amount of data required to transition out aVM150 can be reduced by maintaining a prior snapshot of aVM150 withinsecondary memory144, and transitioning only a “delta” representing a difference of a currently executing version of thatVM150. Such a technique is generally not suitable for reducing the amount of data required to transition aVM150 back into theactive pool148, as a goal of the current technique is to reduce resource usage of theactive pool148. However, in many instance, it may not be necessary for all data of aVM150 to be transitioned fromsecondary memory144 to primary memory in order for theVM150 to execute within theactive pool148. Rather, only a minimum portion of that data (e.g., a state of CPU registers) may be transitioned initially in order to facilitate execution of theVM150. The remaining data may be transitioned “on read” of the data during execution of a task. For example, as aVM150 attempts to read a portion of data in virtualized RAM of theVM150, that portion of data may be moved fromsecondary memory144 to a primary memory of theworker manager140, and passed to theVM150. Using this technique, the cost of transitioning aVM150 fromsecondary memory144 to theactive pool148 may be reduced considerably. For example, embodiments of the present disclosure may assume, for the purposes of calculating a cost of transitioning aVM150 into theactive pool148, only the minimum resources needed to facilitate execution of theVM150. Theworker manager140 may further predict a timing of this “transfer-in” transition as a latest point, before the expected next request, at which the phase-in can occur such that at the time of the expected next request, theVM150 is executing within theactive pool148 and ready to execute the task.

The resource cost during the “idle” phase of a transition tosecondary memory144 can be calculated as the amount ofsecondary memory144 needed to maintain theVM150 in an inactive state during the period until the next expected request, subtracting the timing expected to be required for the transfer in and transfer out phases. This idle time cost can be weighted according to the resources used and combined with weighted representations of the transfer in and transfer out phase costs to result in a calculated cost of transitioning theVM150 intosecondary memory144.

In the instance that theworker manager140 includes multiple tiers ofsecondary memory144, the above-noted costs can be calculated for each available tier ofsecondary memory144. In general, the cost of resources of each lower tier ofsecondary memory144 is expected to be lower (given the expected greater availability of such resources), but the time required to transition in and out aVM150 from such memory is expected to be higher. Thus, larger more frequently usedVMs150 may incur minimal costs when transitioned tohigher tier memories144, while smaller less frequently usedVMs150 may incur minimal costs when transitioned tolower tier memories144.

At (6), themanagement unit142 compares the previously calculated costs, and determines a desired memory state of aVM150 based on a minimal calculated cost. Illustratively, where the cost of keeping aVM150 idle within theactive pool148 is lower than the costs of transitioning theVM150 to asecondary memory144, themanagement unit142 can determine that theVM150 should remain within theactive pool148. For the purposes of description, it is assume that at (6) themanagement unit142 determines that a cost of transitioning theVM150 to asecondary memory144 is less than the cost of keeping theVM150 idle within theactive pool148. Thus, at (6), themanagement unit142 determines that the VM should be transitioned out of theactive pool148 and into secondary memory.

Illustrative interactions for transitioning aVM150 from theactive pool148 to a secondary memory144A, and back into theactive pool148 to service an expected customer request, are depicted withinFIG.4. The interactions ofFIG.4 may illustratively occur subsequent to those ofFIG.4, after themanagement unit142 determines that aVM150 should be transitioned from theactive pool148 to asecondary memory144.

The interactions ofFIG.4 begin at (1), where themanagement unit142 transmits instructions to theactive pool148 to being transitioning aVM150 from theactive pool148 to thesecondary memory144. In the context ofFIG.4, the transmission of instructions may correspond, for example, to software implementing themanagement unit142 transmitting a request to a hypervisor or other “DOM0” or lower-level operating system to initiate snapshotting of theVM150.

At (2), theactive pool148 transitions theVM150 tosecondary memory144. Transitioning of theVM150 can include, for example, storing a snapshot of the state of theVM150 within thesecondary memory144. As noted above, the snapshot may be “cascading” and thus reference a prior snapshot of theVM150 within thesecondary memory144 in order to minimize the data transfer required from a primary memory. Thereafter, theVM150 can remain insecondary memory144 until required (or expected to be required), reducing the overall computing resource usage of theworker manager140.

Thereafter, at (3), themanagement unit142 can detect or predict a call to execute a task within theVM150. Detecting such a call can for example correspond to receiving a request to execute the task, which request may have been transmitted by a user or system external to the on-demandcode execution system110. Predicting a call can for example correspond to detecting that the expected time of a next call (e.g., as used to determine whether to transition theVM150 tosecondary memory144, illustrative interactions for which are described above with reference toFIG.3) is imminently approaching, such that a transition-in phase for theVM150 should begin to ensure that theVM150 is ready to execute the task at the expected time of the next call.

In response to detecting or predicting a next call, themanagement unit142 at (4) instructs theactive pool148 to transition theVM150 from thesecondary memory144 to theactive pool148. In one embodiment, such instructions correspond to instructing a hypervisor or other host operating system to initialize theVM150 based on a snapshot of theVM150 within thesecondary memory144. At (5), theactive pool148 transitions theVM150 to theactive pool148, placing theVM150 in an executing state. Thus, theVM150 can be rendered ready to receive a call to execute the task at the expected time of that call, enabling total resources used on theworker manager140 to be reduced relative to maintaining theVM150 in theactive pool148 in an idle state, while minimally or not impacting responsiveness of theVM150 to requests to execute the task.

With reference toFIG.5 oneillustrative routine500 for managing a memory state of an execution environment, such as aVM150, based on an expected next use of the environment to execute a task on the on-demandcode execution system110 will be described. The routine500 may be carried out, for example, by theworker manager140 ofFIG.1 (e.g., by use of the management unit142).

The routine500 begins atblock502, where theworker manager140 detects an idle environment associated with a task. Detection of an idle environment can correspond, for example, to detecting that an execution of the task has completed, and that no additional requests to execute the task are pending.

The routine500 then continues to block504, where theworker manager140 predicts an idle period for the environment, corresponding to an amount of time until a next request to execute the task within the environment. As discussed above, the idle period may be predicted based on statistical analysis of call history information for a task. Illustratively, the idle period may be calculated from a past distribution of times between calls. For example, the idle period may be calculated such that there is a n % chance that a call is not received before the expiration of the idle period. In some instances, additional statistical analysis, such as a regression analysis, may be applied to determine an expected idle period based on other factors, such as a time-of-day. For example, call history may indicate that a given task is called frequently during business hours but infrequently during nighttime hours. A regression analysis of historical information can thus be used to determine a relationship between current time and a period between calls, which can be applied to a current time (e.g., a time of implementing the routine500) to determine an expected next call to execute a task. In some instances, such statistical analysis may not be required, and other information may be used to predict a time of a next request to execute a task. For example, where thesystem110 is configured to execute a task with a set periodicity (e.g., every minute), the idle period can be calculated from that periodicity, without requiring statistical analysis of historical calls.

The routine500 continues to block506, where theworker manager140 determines an expected resource cost of maintaining the environment in an executing state within primary memory. As noted above, the resource cost may be represented as based on the “resource time” of each resource used to maintain the environment in an idle state (e.g., as a time-factored metric). For example, the RAM usage of an environment may be measured in “megabyte-seconds,” or the amount of megabytes used to maintain the environment in an idle state multiplied by the seconds during which the megabytes would be used (e.g., assuming the predicted idle period is correct). As another example, CPU usage may be measured in CPU cycles, bandwidth may be measured in total data transferred, etc. In one embodiment, each resource time is weighted according to the value or scarcity of the resource and used to calculate an overall (e.g., resource agnostic) cost for maintaining the environment in an idle state.

Similarly, atblock508, theworker manager140 determines an expected resource cost for transitioning the environment to secondary memory during an idle period, and back into primary memory at a time of a request to utilize the environment, or just prior to that request being expected. As with the cost ofblock506, the transition cost may be calculated as a weighted combination of resource time for each resource used to transition the environment. Illustratively, the transition cost account for resources used to transfer out an environment from primary memory, maintain the environment in secondary memory, and transfer in the environment back into primary memory to execute code within the environment.Block508 may be repeated for each potential secondary memory. As noted above, because the speed of each tier of secondary memory is expected to vary, the costs of transferring an environment to each tier is also expected to vary. However, these costs may not vary with any particular aspect of an environment, but rather with a combination of numerous factors, including the amount of data needed to be transferred for an environment (e.g., the “working set”) and the expected time until next use of that environment.

Atblock510 of the routine500, theworker manager140 conducts a comparison of the costs calculated in the above-notedblocks506 and508, and determines whether the cost to transition the environment to a secondary memory is less than the cost of maintaining the environment idle within primary memory. Ifblock510 evaluates to false, the routine500 proceeds to block518 and ends. This scenario illustratively corresponds to an instance in which a next call to utilize the environment is expected to come relatively quickly, and/or where the cost of transitioning the environment to a secondary memory is very high (e.g., the environment has a large working set of memory). In such an instance, maintaining the environment within primary memory may represent the most efficient use of resources on thesystem110.

Alternatively, ifblock510 evaluates to true, the routine500 continues to block512, where theworker manager140 transitions the environment to the secondary memory associated with a lowest expected cost. Where the environment is a virtual machine, such a transition can include snapshotting the environment and storing the snapshot in the secondary memory. Where the environment is a software container or other software construct, such a transition can include checkpointing or “freezing” the environment and storing the checkpoint within the secondary memory.

Thereafter, atblock514, theworker manager140 detects or predicts a call to execute the task associated with the environment. Detecting the call can correspond to receiving an instruction or request to execute the task. Predicting a call can correspond to determining that the predicted idle period of the environment (e.g., as determined at block504) is nearing completion, such that a transition of the environment into primary memory should begin to enable the environment to be executing at the time when the predicted idle period ends. Thus, atblock516, theworker manager140 transitions the environment back into primary memory, and sets the environment to executing within theactive pool148. Where the environment is a virtual machine, transitioning the environment may include recreating the virtual machine from the snapshot. Where the environment is a software container or other construct, transitioning the environment may include restoring the container or construct from a checkpoint. In either instance, transitioning the environment may rely on functionality to transition data of the environment on an as needed basis, minimizing the amount of data that must be transmitted between the secondary memory and the primary memory to place the environment in a working state. The environment can then facilitate execution of a task within the environment.

Thus, by implementation of the routine500, the overall computing resource usage of thesystem110 with respect to an environment during an idle period can be reduced, while maintaining the ability of thesystem110 to rapidly service requests to execute code within the environment. Indeed, where a subsequent request to utilize an environment is received at or after the expected next use of the environment, the total time of thesystem110 to begin execution of a task is expected to be unchanged. However, because at least a portion of the idle time of the environment is spend stored in secondary memory, primary memory of thesystem110 can be freed during that idle time to facilitate other executions of tasks. Thus, the overall capacity of thesystem110 to execute tasks can be increased.

With reference toFIGS.6A and6B, illustrative interactions will be described for gracefully recovering network connections after transitioning of aVM instance150 between state or host devices. Specifically,FIG.6A depicts illustrative interactions undertaken on afirst worker manager140A, representing at least one first host device, to notify external devices that each stateful network connection to aVM instance150A is terminated.FIG.6B depicts illustrative interactions to migrate theVM instance150A to asecond worker manager140B representing at least a second host device, and to notify theVM instance150A that each stateful network connection has been terminated, enabling theVM instance150A to reestablish such stateful network connections. While the interactions ofFIG.6A and6B are depicted as occurring in conjunction with migration of aVM instance150 to a different host device, similar interactions could be undertaken on a single host device to facilitate transition of theVM instance150 between memories (e.g., to facilitate halting theVM instance150 for a substantial period of time likely to cause timeouts).

The interactions ofFIG.6A begin at (1), where themanagement unit142 determines that theVM instance150A is to be transitioned. In some instances, themanagement unit142 may determine that such a transition is to occur based on the interactions described above (e.g., based on determined that a next expected use of theinstance150 exceeds a given value). In other instances, themanagement unit142 may determine that such a transition is to occur based on receiving instructions from another device or component. For example, themanagement unit142 may receive instructions from an administrator or control plane device that theinstance150 should be migrated to anotherworker manager140.

At (2), theVM instance150A retrieves from aconnection state store602 of theworker manager140A information identifying current stateful network connections of theVM instance150. Theconnection state store602 may be implemented within memory of theworker manager140A (e.g., a primary memory and/or secondary memory144), and reflect network information used by theworker manager140A to handle network traffic to and from theVM instance150A. Illustratively, theconnection state store602 may reflect a “state table” or other table maintained by a hypervisor, “DOM-0,” lower-level operating system, virtual switch, or the like, which reflects how traffic is routed from a physical network connection of a host device to various execution environments on the host device. Because the host device maintains that physical network connection, it may necessary operate as a middle-man between eachVM instance150 and a physical network, thus enabling the host device to maintain knowledge of network connections of eachVM instance150 to external devices (e.g., external to the VM instance150). Thus, by inquiry to theconnection state store602, themanagement unit142 can receive information identifying open connections of theVM instance150A. In some embodiments, this information may be received independent of communications with theVM instance150A.

Thereafter, at (3), themanagement unit142 transmits, for each open connection of theVM instance150A, a termination signal to the other party of the connection. Illustratively, if theVM instance150A has an active TCP connection to a givenauxiliary service106, themanagement unit142 may transmit a termination signal (e.g., a TCP reset packet) to theauxiliary service106. Illustratively, themanagement unit142 may instruct a hypervisor or other low-level operating system to transmit the signal, or may generate and transmit the signal itself. In one embodiment, themanagement unit142 “masquerades” as theVM instance150A in order to send the termination signal, such as by altering aspects of the termination signal to match what would be expected from such a signal if generated at theVM instance150A. Illustratively, themanagement unit142 may alter a source network address to match an address of theVM instance150A, or modify a sequence number to match a current sequence number for the connection (e.g., as indicated in the connection state store602). Thus, each other party to a stateful network connection of theVM instance150A will understand that such connection has terminated.

In one embodiment, the interactions ofFIG.6A may occur without notification to theVM instance150A. Thus, from the point of view of theVM instance150A, each network connection will be understood to remain open. Beneficially, this may prevent theVM instance150A from attempting to reestablish the connections. In some instances, themanagement unit142 may be configured to prevent theVM instance150A from receiving further communications from an external device after sending a termination signal to the external device, to prevent the external device from prematurely reestablishing the network connection. Thus, via the interactions ofFIG.6A, each external device with a stateful network connection to aVM instance150A may be notified that such connection has terminated.

The interactions ofFIG.6A are continued inFIG.6B, which depicts illustrative interactions for migrating aVM instance150A to aworker manager140B, and for notifying theVM instance150A that stateful network connections to external devices have been terminated, in order to enable theinstance150A to gracefully recover those connections. The interactions ofFIG.6B illustratively occur subsequently to those ofFIG.6A, and the numbering ofFIG.6A is therefore continued inFIG.6B. However, the interactions of these figures may in some embodiments be implemented separately.

With reference toFIG.6B, at (4), theVM instance150A is migrated from theworker manager140A to theworker manager140B. A number of mechanisms for virtual machine migration are known in the art, and therefore will not be discussed in detail herein. Any such suitable migration may be utilized in accordance with embodiments of the present disclosure. Additionally or alternatively, migration of aVM instance150A may include transfer of a representation of theVM instance150A, such as a snapshot, from a memory of theworker manager140A to a memory of theworker manager140B. In some instances, transfer of such representation may include transfer of a “delta” between a snapshot of theVM instance150A on theworker manager140A and a related snapshot maintained in a memory ofworker manager140B. As noted above, various snapshots in the on-demandcode execution system110 may be cascading or interrelated. Thus, where theVM instance150A was generated based on a snapshot also maintained at theworker manager140B (or related to a snapshot maintained at theworker manager140B), migration of theVM instance150A may require transfer to theworker manager140B of only the differences between the snapshot of theVM instance150A on theworker manager140A and a related snapshot on theworker manager140B. In some instances, related snapshots (e.g., representing “ancestors” states of various VM instances150) may be distributed amongworker managers140 in thesystem110 to facilitate later rapid migration ofVM instances150, if required.

In addition to information representing theVM instance150A, theworker manager140A transfers to theworker manager140B, at (4), information representing stateful network connections understood by theVM instance150A to have been maintained at the time that theinstance150A was halted. The information correspond, for example, to that retrieved from theconnection state store602 at interaction (2) ofFIG.6A, above.

Thereafter, at (5), themanagement unit142 determines that theVM instance150A should be placed into theactive pool148B on theworker manager140B. The decision to place theVM instance150 into theactive pool148B may be based, for example, on a predicted request to execute a task within theinstance150A.

As noted above, theVM instance150A is assumed to have been halted at a point where it understands one or more stateful network connections to be active. However, migration of theVM instance150A may impair such connections, for example because an external device understands a timeout to have occurred on the connection, or because a change in location of theVM instance150A has changed a parameter of the connection (such as a source IP address). To reduce the likelihood of errors occurring due to theVM instance150A attempting to use impaired stateful connections, themanagement unit142, at (6), transmits to theVM instance150A termination signals for each stateful network connections that theVM instance150A understands to be active (e.g., as indicated in the connection state information received from theworker manager140A). Transmission may include, for example, causing a host device for theVM instance150A to generate and transmit termination signals to theVM instance150A. A termination signal may correspond, for example, to a TCP reset packet. In one embodiment, the termination signals may be generated so as to appear to originate from the external devices to which the network connections were made. For example, themanagement unit142 may modify a source IP address of each termination signal to match an IP address of the relevant external device, and may include within the termination signal a next sequence number for the connection. Thus, theVM instance150A can be expected to process the termination signal as if it originated from the external device.

The termination signals may illustratively be transmitted during initialization of theVM instance150A, before a task execution begins. Alternatively, termination signals may be transmitted after execution of code begins. Prior to transmission of termination signals, themanagement unit142 may in some instances inhibit communication between theVM instance150A and external devices to which the network connections were made, to prevent errors related to attempted use of such connections.

Because both theVM instance150A and the relevant external devices have received termination signals related to stateful network connections, these endpoints can be expected to have shared, synchronized knowledge of the state of those network connections. As such, at (7), theVM instance150A can gracefully reestablish network connections required for further processing on theVM instance150A. Illustratively, theVM instance150A may be configured to execute code that detects termination of stateful network connections, and attempts to reestablish those connections. As such, on reception of the termination signals from themanagement unit142, theVM instance150A can communicate with the relevant external device (e.g., auxiliary services106) to reestablish stateful network connections.

While the interactions ofFIG.6A and6B depict notifying both parties to a stateful network connection of termination of that connection, some embodiments of the present disclosure may notify only one such party. For example, where aVM instance150 is the initiator of such a connection, it may be sufficient to notify only theVM instance150 that the connection has terminated to cause theVM instance150 to reestablish the connection. However, notification of the external device as well may in some instance mitigate problems such as an abundance of open connections at the external device, which might otherwise prohibit reestablishment of a connection to theVM instance150 at a later time or from another location.

Oneillustrative routine700 for ending stateful connections of aVM instance150 in connection with transitioning of theVM instance150 is depicted inFIG.7. The routine700 may be implemented in one or more contexts, identified inFIG.7 as Context A and Context B. In one embodiment, Context A and B correspond to different host devices. Thus, routine700 may be implemented when migrating theVM instance150. In another embodiment, Context A and B are the same host device. Thus, routine700 may be implemented when transitioning theVM instance150 between memory states (e.g., when theVM instance150 is expected or could potentially be inactive for a period of time that causes stateful network connections to be broken).

The routine700 may be carried out, for example, by one or more management units142 (e.g., by controlling a host device hosting a VM instance150). Illustratively, amanagement unit142 of a first host device representing Context A may implementblocks702 through706 of the routine700, while amanagement unit142 of a second host device representing Context B may implement blocks710 through714 of the routine. Themanagement units142 of each context may cooperate to implementblock708, which as discussed below represents transition of theVM instance150 between contexts. Where Context A and B are the same context, one ormore management units142 of that context may implement the routine700 in its entirety.

The routine700 begins atblock702, where themanagement unit142 determines that theVM instance150A is to be transitioned. In some instances, themanagement unit142 may determine that such a transition is to occur based on the interactions described above (e.g., based on determined that a next expected use of theinstance150 exceeds a given value). In other instances, themanagement unit142 may determine that such a transition is to occur based on receiving instructions from another device or component. For example, themanagement unit142 may receive instructions from an administrator or control plane device that theinstance150 should be migrated to anotherworker manager140.

Atblock704, themanagement unit142 identifies stateful network connections maintained by theVM instance150. The stateful network connections may be determined, for example, by interrogation of a host device hosting theVM instance150, which may maintain information related to open network connections of theVM instance150 in order to facilitate routing of network data packets to theVM instance150.

Atblock706, themanagement unit142, for each open network connection, transmits to an external device that is party to the connection a termination signal. The termination signal may illustrative correspond to a TCP reset data packet (e.g., a TCP formatted packet with a reset bit flag set to a true value). In one embodiment, themanagement unit142 “masquerades” as theVM instance150 in order to send the termination signal, such as by altering aspects of the termination signal to match what would be expected from such a signal if generated at theVM instance150. Illustratively, themanagement unit142 may alter a source network address to match an address of theVM instance150, or modify a sequence number to match a current sequence number for the connection (e.g., as indicated in the connection state store602). Thus, the other parties to the stateful network connections of theVM instance150 will understand that such connection has terminated.

Atblock708, themanagement unit142 transitions theVM instance150 from Context A to Context B. In one embodiment, implementation ofblock708 may include migration of theVM instance150 from a first host device to a second host device. In another embodiment, implementation ofblock708 may include transitioning theVM instance150 from an active state to an inactive state, and back to an active state at a later point in time (e.g., to save resources associated with maintaining theVM instance150 in an active state).

Atblock712, themanagement unit142 again determines open network connections that theVM instance150 understands to be open. Where Context A and Context B are the same host device, block712 may be duplicative and thus omitted. Where Context A and Context B are different host devices, block712 may include amanagement unit142 of a host device corresponding to Context B receiving a listing of open connections from amanagement unit142 of a host device corresponding to Context A.

Atblock714, themanagement unit142 transmits to theVM instance150 termination signals for each stateful network connections that theVM instance150 understands to be open. In one embodiment, themanagement unit142 may generate the termination signals so as to appear to originate from the external devices to which the network connections were made. For example, themanagement unit142 may modify a source IP address of each termination signal to match an IP address of the relevant external device, and may include within the termination signal a next sequence number for the connection. Thus, theVM instance150 can be expected to process the termination signal as if it originated from the external device.

The routine700 then ends atblock716. As theVM instance150 may be assumed to be executing code that operates to reestablish network connections on reception of a termination signal for those connections (as necessary), implementation of the routine700 can thus provoke theVM instance150 into reestablishing those connections. Moreover, because the routine700 notifies both an external device and theVM instance150 of termination of network connections, the parties to the connection are unlikely to attempt to utilize the connection during transition of theVM instance150, which may be beneficial as such use may result in errors.

As will be appreciated by one skilled in the art, the routine700 may in some embodiments be implemented by a host device hosting aVM instance150, without requiring control ofVM instance150 or external devices to whichVM instances150 communicate. Such operation may be desirable to reduce or eliminate the need to customize code executing onVM instances150 or external devices.

All of the methods and processes described above may be embodied in, and fully automated via, software code modules executed by one or more computers or processors. The code modules may be stored in any type of non-transitory computer-readable medium or other computer storage device. Some or all of the methods may alternatively be embodied in specialized computer hardware.

Conditional language such as, among others, “can,” “could,” “might” or “may,” unless specifically stated otherwise, are otherwise understood within the context as used in general to present that certain embodiments include, while other embodiments do not include, certain features, elements and/or steps. Thus, such conditional language is not generally intended to imply that features, elements and/or steps are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without user input or prompting, whether these features, elements and/or steps are included or are to be performed in any particular embodiment.

Disjunctive language such as the phrase “at least one of X, Y or Z,” unless specifically stated otherwise, is otherwise understood with the context as used in general to present that an item, term, etc., may be either X, Y or Z, or any combination thereof (e.g., X, Y and/or Z). Thus, such disjunctive language is not generally intended to, and should not, imply that certain embodiments require at least one of X, at least one of Y or at least one of Z to each be present.

Unless otherwise explicitly stated, articles such as ‘a’ or ‘an’ should generally be interpreted to include one or more described items. Accordingly, phrases such as “a device configured to” are intended to include one or more recited devices. Such one or more recited devices can also be collectively configured to carry out the stated recitations. For example, “a processor configured to carry out recitations A, B and C” can include a first processor configured to carry out recitation A working in conjunction with a second processor configured to carry out recitations B and C.

Any routine descriptions, elements or blocks in the flow diagrams described herein and/or depicted in the attached figures should be understood as potentially representing modules, segments, or portions of code which include one or more executable instructions for implementing specific logical functions or elements in the routine. Alternate implementations are included within the scope of the embodiments described herein in which elements or functions may be deleted, or executed out of order from that shown or discussed, including substantially synchronously or in reverse order, depending on the functionality involved as would be understood by those skilled in the art.

It should be emphasized that many variations and modifications may be made to the above-described embodiments, the elements of which are to be understood as being among other acceptable examples. All such modifications and variations are intended to be included herein within the scope of this disclosure and protected by the following claims.

Claims (20)

What is claimed is:

1. A system to facilitate reconnection of stateful network connections after migration of a virtual machine instance, the system comprising:

a first host computing device and a second host computing device;

wherein the first host computing device is configured with computer-executable instructions to:

host the virtual machine instance prior to the migration;

identify a stateful network connection between the virtual machine instance and an external device external to the virtual machine instance;

independent of communications of the external device and virtual machine instance, transmit to the external device a first termination signal indicating termination of the stateful network connection, wherein the first termination signal is formatted to indicate that the first termination signal originates from the virtual machine instance; and

migrate the virtual machine instance to the second host computing device;

wherein the second host computing device is configured with computer-executable instructions to:

host the virtual machine instance subsequent to the migration;

identify the stateful network connection; and

independent of communications of the external device and virtual machine instance, transmit to the virtual machine instance a second termination signal indicating termination of the stateful network connection, wherein the second termination signal is formatted to indicate that second termination signal originates from the external device; and

wherein the virtual machine instance is configured, on receiving the second termination signal, to attempt to reestablish the stateful network connection.

2. The system ofclaim 1, wherein the stateful network connection is a network connection utilizing the transport control protocol (TCP).

3. The system ofclaim 1, wherein the first host computing device is configured with the computer-executable instructions to identify the stateful network connection based at least in part on state information generated at the first host computing device in response to network data transmitted from the virtual machine instance to the external device.

4. The system ofclaim 1, wherein the first host computing device is configured with the computer-executable instructions to format the second termination signal to indicate that it originates from the external device at least partly by including within the second termination signal a network address of the external device.

5. The system ofclaim 1, wherein the first host computing device is configured with the computer-executable instructions to format the first termination signal to indicate that it originates from the virtual machine instance at least partly by including within the first termination signal a next sequence number associated with the stateful network connection.

6. A computer-implemented method implemented by one or more host computing devices, the computer-implemented method comprising:

identifying a stateful network connection between a virtual machine instance hosted by the one or more host computing devices and an external device external to the virtual machine instance;

independent of communications of the external device and virtual machine instance, transmitting to the external device a first termination signal indicating termination of the stateful network connection;

temporarily halting execution of the virtual machine instance; and

in response to resuming execution of the virtual machine instance on a host computing device, and independent of communications of the external device and virtual machine instance, transmitting, to the virtual machine instance executing on the host computing device, a second termination signal indicating termination of the stateful network connection;

wherein the virtual machine instance is configured, on receiving the second termination signal, to attempt to reestablish the stateful network connection.

7. The computer-implemented method ofclaim 6, wherein the one or more host computing devices comprise a single host computing device, and wherein temporarily halting execution of the virtual machine instance comprises transitioning of the virtual machine instance to secondary memory of the single host computing device.

8. The computer-implemented method ofclaim 6, wherein the one or more host computing devices comprise a first host computing device and second host computing device, and wherein temporarily halting execution of the virtual machine instance comprises migrating the virtual machine instance from the first host computing device to the second host computing device.

9. The computer-implemented method ofclaim 8, wherein migrating the virtual machine instance from the first host computing device to the second host computing device comprises generating a snapshot of the virtual machine instance at the first host computing device and transmitting the snapshot to the second host computing device.

10. The computer-implemented method ofclaim 9, the snapshot of the virtual machine instance represents a difference in state of the virtual machine instance from a prior snapshot of the virtual machine instance maintained at the second host computing device.

11. The computer-implemented method ofclaim 6, wherein the first termination signal is formatted to indicate that the first termination signal originates from the virtual machine instance.

12. The computer-implemented method ofclaim 6, wherein the first termination signal is formatted to indicate that the first termination signal originates from the virtual machine instance at least partly by inclusion in the first termination signal of a next sequence number for the stateful network connection.

13. The computer-implemented method ofclaim 12, transmitting to the external device the first termination signal indicating termination of the stateful network connection occurs subsequent to resuming execution of the virtual machine instance.

14. The computer-implemented method ofclaim 12, wherein identifying the stateful network connection comprises identifying a plurality of stateful network connections, and wherein transmitting to the virtual machine instance the second termination signal indicating termination of the stateful network connection comprises transmitting to the virtual machine instance a plurality of second termination signals, each of the second termination signals indicating termination of a stateful network connection of the plurality of stateful network connections.

15. Non-transitory computer-readable media comprising instructions executable by a computing system to:

identify a stateful network connection between a virtual machine instance hosted by the computing system and an external device external to the virtual machine instance;

temporarily halt execution of the virtual machine instance; and

in response to resuming execution of the virtual machine instance on a host computing device, and independent of communications of the external device and virtual machine instance, transmit, to the virtual machine instance executing on the host computing device, a termination signal indicating termination of the stateful network connection;

wherein the virtual machine instance is configured, on receiving the termination signal, to attempt to reestablish the stateful network connection.

16. The non-transitory computer-readable media ofclaim 15, wherein the termination signal is formatted to indicate that the termination signal originates from the external device.

17. The non-transitory computer-readable media ofclaim 15, wherein the termination signal is formatted to indicate that the termination signal originates from the external device at least partly by including within the termination signal a next sequence number associated with the stateful network connection.

18. The non-transitory computer-readable media ofclaim 15 further comprising instructions executable by the computing system to transmit to the external device a second termination signal indicating termination of the stateful network connection, wherein the second termination signal is transmitted independent of communications of the external device and virtual machine instance.

19. The non-transitory computer-readable media ofclaim 18, wherein the second termination signal is formatted to indicate that the second termination signal originates from the virtual machine instance.

20. The non-transitory computer-readable media ofclaim 15, wherein the stateful network connection comprises a plurality of stateful network connections, and wherein the instructions are executable by the computing system to transmit to the virtual machine instance a plurality of second termination signals, each of the second termination signals indicating termination of a stateful network connection of the plurality of stateful network connections.

Images